Download the long-term evolution of multilocus traits under frequency

Evolution, 60(11), 2006, pp. 2226–2238
THE LONG-TERM EVOLUTION OF MULTILOCUS TRAITS UNDER
FREQUENCY-DEPENDENT DISRUPTIVE SELECTION
G. SANDER
VAN
DOORN1,2
AND
ULF DIECKMANN3,4
1 Centre
for Ecological and Evolutionary Studies, University of Groningen, Kerklaan 30, 9751 NN Haren, The Netherlands
and Ecology Program, International Institute for Applied Systems Analysis, Schlossplatz 1, 2361 Laxenburg, Austria
4 Section Theoretical Biology, Institute of Biology, Leiden University, Kaiserstraat 63, 2311 GP Leiden, The Netherlands
3 Evolution
Abstract. Frequency-dependent disruptive selection is widely recognized as an important source of genetic variation.
Its evolutionary consequences have been extensively studied using phenotypic evolutionary models, based on quantitative genetics, game theory, or adaptive dynamics. However, the genetic assumptions underlying these approaches
are highly idealized and, even worse, predict different consequences of frequency-dependent disruptive selection.
Population genetic models, by contrast, enable genotypic evolutionary models, but traditionally assume constant fitness
values. Only a minority of these models thus addresses frequency-dependent selection, and only a few of these do so
in a multilocus context. An inherent limitation of these remaining studies is that they only investigate the short-term
maintenance of genetic variation. Consequently, the long-term evolution of multilocus characters under frequencydependent disruptive selection remains poorly understood. We aim to bridge this gap between phenotypic and genotypic
models by studying a multilocus version of Levene’s soft-selection model. Individual-based simulations and deterministic approximations based on adaptive dynamics theory provide insights into the underlying evolutionary dynamics.
Our analysis uncovers a general pattern of polymorphism formation and collapse, likely to apply to a wide variety
of genetic systems: after convergence to a fitness minimum and the subsequent establishment of genetic polymorphism
at multiple loci, genetic variation becomes increasingly concentrated on a few loci, until eventually only a single
polymorphic locus remains. This evolutionary process combines features observed in quantitative genetics and adaptive
dynamics models, and it can be explained as a consequence of changes in the selection regime that are inherent to
frequency-dependent disruptive selection. Our findings demonstrate that the potential of frequency-dependent disruptive
selection to maintain polygenic variation is considerably smaller than previously expected.
Key words. Adaptive dynamics, evolutionary branching, Levene model, maintenance of genetic variation, population
genetics, protected polymorphism, symmetry breaking.
Received May 16, 2006.
Frequency-dependent selection plays an important role in
the origin and maintenance of genetic variation (Felsenstein
1976; Slatkin 1979; Hedrick et al. 1976). Conditions for stable polymorphisms are much relaxed when fitness values are
not constant but vary with the frequency of different genotypes present in a population. Protected polymorphisms can
then be established whenever rare genotypes have a selective
advantage (Lewontin 1958). This may even lead to situations
in which, at population genetic equilibrium, the heterozygote
resulting from an allelic dimorphism experiences a fitness
disadvantage. (Note that this is the exact opposite of the
situation required for stable polymorphisms to occur with
constant fitness values.) In such a case, the population is
caught at a fitness minimum, at which it experiences disruptive selection.
The consequences of such frequency-dependent disruptive
selection have most extensively been investigated in the context of quantitative genetics (e.g., Slatkin 1979; Bulmer 1980)
and in the related frameworks of evolutionary game theory
(e.g., Maynard Smith 1982; Hofbauer and Sigmund 1998)
and adaptive dynamics (e.g., Dieckmann and Law 1996; Metz
et al. 1996; Geritz et al. 1998; Hofbauer and Sigmund 1998).
See Abrams (2001) for a comparison of these three methods.
Although the insights obtained through these different approaches are similar in some respects (Taylor 1996), their
predictions for the effects of frequency-dependent disruptive
2 Present address: Section of Integrative Biology, University of
Texas, 1 University Station C0930, Austin, Texas 78712; E-mail:
[email protected].
Accepted July 10, 2006.
selection are strikingly different. In quantitative genetics
models, the maintenance of genetic variation results from the
broadening of continuous phenotypic distributions exposed
to such selection. In adaptive dynamics models, frequencydependent disruptive selection can cause evolutionary
branching (Metz et al. 1996; Geritz et al. 1997, 1998). Such
branching processes characteristically involve the convergence of a monomorphic population to a fitness minimum,
followed by the adaptive emergence and further diversification of a discrete polymorphism.
The discordance of these predictions is caused by the different genetic assumptions underlying quantitative genetics
and adaptive dynamics models. Quantitative genetics models
are often purely phenomenological, but when a mechanistic
underpinning is given, it is usually assumed that phenotypic
characters are influenced by a large number of loci, each of
which contributes only marginally to the phenotype. In every
generation, the genetic variation present in the parent generation is redistributed among the offspring through recombination and segregation, that is, as a consequence of sexual
reproduction. Because many loci are involved in this process,
the distribution of phenotypes in the population is continuous
and normal. Adaptive dynamics models, in contrast, usually
consider asexual reproduction (or single-locus, haploid genetics) and monomorphic populations (for exceptions see
Kisdi and Geritz 1999; Van Dooren 1999).
From the viewpoint of population genetics, the assumptions of infinite loci with infinitesimal effects (quantitative
genetics) or of asexual reproduction (adaptive dynamics) are
both highly idealized. It is therefore difficult to predict the
2226
䉷 2006 The Society for the Study of Evolution. All rights reserved.
2227
EVOLUTIONARY BRANCHING OF MULTILOCUS TRAITS
effect of frequency-dependent disruptive selection for realistic genetic settings. Despite the fact that frequency-dependent selection has been included in the theory of population
genetics right from its conception (Fisher, 1930), most of
population genetics theory assumes constant fitness values
(for exceptions see Clarke 1972; Cockerham et al. 1972;
Cressman 1992); such theory cannot be used to predict the
consequences of frequency-dependent selection. In particular, the evolutionary dynamics of multilocus characters under
frequency-dependent disruptive selection remains elusive.
However, several attempts have been made to bridge the gap
between population genetic and phenotypic models of frequency-dependent selection, and particularly the integration
of population genetics with evolutionary game theory has
received considerable attention (e.g., Cressman, 1992; Hofbauer and Sigmund 1998).
As a case in point, Bürger (2002a,b) presented a population
genetic analysis of a model of intraspecific competition that
had previously been analyzed within both the quantitative
genetics (e.g., Slatkin 1979) and the adaptive dynamics
framework (e.g., Metz et al. 1996). Bürger focused on the
dynamics and population genetic equilibria of the frequencies
of a fixed set of alleles in a multilocus model with frequencydependent disruptive selection. He investigated the conditions under which disruptive selection on the phenotypes can
be observed and quantified the amount of genetic variation
that can be maintained. The analyzed model exhibits a number of unexpected phenomena, which underscore that the population genetics of frequency-dependent disruptive selection
can be surprisingly complex.
A complementary approach was initiated by Kisdi and Geritz (1999) and Van Dooren (1999), who extended adaptive
dynamics models by incorporating diploid genetics and sexual reproduction. Focusing on the simplest model of interest,
these authors studied the evolution of alleles at a single locus
under frequency-dependent disruptive selection. Unlike models that consider a fixed and limited set of alleles, these analyses explicitly considered mutations with small phenotypic
effects. Long-term evolution can then proceed as a sequence
of substitution steps during which existing alleles are replaced by novel ones created by mutation. Similar approaches
have been developed in the population genetics literature
(e.g., Keightley and Hill 1983; Bürger et al. 1989). As in
asexual adaptive dynamics models, frequency-dependent disruptive selection can cause evolutionary branching in diploid
sexual populations (this occurs under the same conditions as
in asexual models), leading to the establishment of a polymorphism of alleles (Kisdi and Geritz 1999). As a consequence of the constraints imposed by random mating and
segregation, the evolution of dominance-recessivity relations
between the alleles is selectively favored (Van Dooren 1999).
In this paper, we aim to extend the understanding of the
long-term consequences of frequency-dependent disruptive
selection by analyzing mutations and allele substitutions in
a multilocus model. This approach extends the earlier work
of Bürger (2002a,b) by allowing for long-term evolution by
mutations and allele substitutions. At the same time, our work
extends the analysis by Kisdi and Geritz (1999) by allowing
for multilocus genetics. We will consider Levene’s soft-selection model (Levene 1953) as a prototypical example of
situations generating frequency-dependent disruptive selection. Levene’s model is commonly used for studying the
maintenance of variation in a heterogeneous environment; it
is relatively simple, its population genetics are well known
(Roughgarden 1979), and it has been considered in several
related studies (Kisdi and Geritz 1999; Van Dooren 1999;
Spichtig and Kawecki 2004).
MODEL DESCRIPTION
Ecological Assumptions
We consider an organism with discrete, nonoverlapping
generations in a heterogeneous environment consisting of two
habitats. Individuals are distributed at random over the two
habitats at the start of each generation. The two habitats differ
in ecological conditions such that an individual is more or
less adapted to a habitat depending on its ecological strategy
z, a one-dimensional quantitative character. Specifically, we
assume that an individual’s viability in habitat i ⫽ 1, 2 is
given by
[
]
1
v i (z) ⫽ exp ⫺ (z ⫺ ␮ i ) 2 /␴ 2 ,
2
(1)
which implies that the optimal phenotype is ␮1 in the first
habitat and ␮2 in the second. The parameter ␴ is an inverse
measure of the intensity of local selection and determines
how rapidly viability declines with the difference between
an individual’s ecological strategy and the locally optimal
one. Without loss of generality, we set ␮1 ⫽ ⫺␮2 ⫽ ␮.
We assume soft selection (sensu Levene 1953; see also
Ravigné et al. 2004); in each generation, a fixed number fi N
of randomly chosen adults are recruited from habitat i;
throughout, we set f1 ⫽ f2 ⫽ ½. These adults form a single
mating population of population size N, in which mating
occurs at random and offspring are produced at the end of
each generation.
Genetic Assumptions
The ecological strategy z is encoded by L diploid loci. One
or more distinct alleles may occur at every locus. We use
the index k to arbitrarily label the different alleles that occur
within the population at a specified locus l. Correspondingly,
alk denotes the kth allele at the lth locus, and xlk denotes its
phenotypic effect (allelic effect). We initially assume that
loci are unlinked and that alleles interact additively at each
locus and between loci. Hence, for an individual carrying
alleles alk⬘ and alk⬙ at the lth locus, the phenotypic effect of
this locus is given by yl ⫽ xlk⬘ ⫹ xlk⬙, and the individual’s
ecological strategy is given by
冘y.
L
z⫽
l⫽1
l
(2)
Later in this study, we will also consider nonadditive interactions within and between loci, as well as genetic linkage
between loci.
Unlike previous models (reviewed in Felsenstein 1976;
Hedrick et al. 1976), which were concerned with the shortterm evolutionary process of changes in allele frequencies,
we do not constrain the set of alleles that may be present in
2228
G. S.
VAN
DOORN AND U. DIECKMANN
the population. By allowing new alleles to appear through
mutation, we can study the long-term evolutionary process
of changes in the phenotypic effects of alleles. Mutations
occur at rate m per allele per generation and change the phenotypic effect of an allele by an amount drawn from a normal
distribution with zero mean and standard deviation ␴m.
Simulation Details
In addition to analytic approximations, we consider an individual-based model. At the start of a simulation, the population is initialized by creating N identical individuals that
are homozygous at all loci. Although the initial population
exhibits no genetic variation, it usually takes just a few generations before mutation has created sufficient genetic variation for adaptive change to occur. The phenotypic effects
of the alleles used to initialize the population is chosen such
that the initial phenotype is far away from the average of the
optimal phenotypes in the two patches. Other choices of the
initial phenotype do not lead to different results: the population always first converges on the phenotype z ⫽ 0, the
average of the patch optima, before further evolution occurs
(see below).
At regular intervals, we determined the distribution of phenotypes in the population and the distributions of the phenotypic effects of alleles at individual loci.
Environmental Variation
Throughout this paper and as indicated by equation (2),
we maintain conceptual simplicity by supposing that the ecological strategy is completely genetically determined. However, effects of the microenvironment on the phenotype can
be incorporated into the model straightforwardly by modifying the selection parameter ␴2 (Bürger 2000, pp. 158–160);
it changes to ␴2 ⫹ Ve, where Ve denotes the variance of the
environmental component of phenotypic variation. The validity of this rescaling argument was confirmed by individualbased simulations (see Figure S1 in the supplementary material available online only at http://dx.doi.org/10.1554/
06-291.1.s1).
It should be noted that variation in the environment itself,
caused by, for instance, variability of the parameters ␮ and
␴, cannot be dealt with simply by rescaling the model. For
this reason, we have explored a small number of scenarios
by means of individual-based simulations (for an example,
see Figure S2 in the supplementary material available online);
these simulations confirmed the robustness of our conclusions. A more comprehensive analysis of the effects of environmental variability on the maintenance of polygenic variation in the model analyzed here is suggested as an interesting topic of future research.
INDIVIDUAL-BASED MODEL
Two Selection Regimes
Our investigations of the individual-based model defined
above show that, not unexpectedly, evolutionary outcomes
critically depend on the relative magnitude of the parameters
␮ and ␴.
When the optimal strategies in the two habitats are not too
different, or when viability selection is weak (␮ ⬍ ␴), longterm evolution of the ecological strategy z proceeds toward
the generalist strategy z* ⫽ 0 (results not shown). Once the
population has reached this generalist strategy, no further
phenotypic evolution takes place. Mutation-selection balance
maintains only a tiny amount of variation in the population.
These observations agree with analytical results (Geritz et al.
1998; Kisdi and Geritz 1999) that predict the strategy z* ⫽
0 to be both convergence stable and evolutionarily stable for
␮ ⬍ ␴. The former implies that evolution through small phenotypic steps will proceed toward z* ⫽ 0, with each step
corresponding to the mutation and subsequent substitution of
an allele. The latter implies that no allele coding for an alternative phenotype will be able to invade once the phenotype
z* ⫽ 0 has been established, and, therefore, that the population experiences stabilizing selection at z* ⫽ 0.
By contrast, when the difference between the optimal strategies is large, or when viability selection is strong (␮ ⬎ ␴),
we observe the emergence of a stable phenotypic polymorphism through the process of evolutionary branching (Metz
et al. 1996; Geritz et al. 1997, 1998). Figure 1 shows a simulation for ␮ ⫽ 1.5 and ␴ ⫽ 1.0 (we performed simulations
for 0.2 ⱕ ␮/␴ ⱕ 2.0 in steps of 0.2 and 2.0 ⱕ ␮/␴ ⱕ 5.0 in
steps of 0.5). Other parameters are set to N ⫽ 1000, L ⫽ 3
(we also considered L ⫽ 1, . . . , 5, 10, 20, 50, 100, 250),
␴2m ⫽ 10⫺3, and m ⫽ 10⫺4; unless stated otherwise, these
parameter values will be used throughout this paper. As illustrated in the left panel of Figure 1, directional evolution
first converges toward the generalist strategy z* ⫽ 0, where
selection turns disruptive. This is because the strategy z* ⫽
0 is convergence stable, but not evolutionarily stable (Geritz
et al. 1998; Kisdi and Geritz 1999). Therefore, alleles coding
for alternative phenotypes can invade the generalist population, thus establishing genetic and phenotypic polymorphism (middle and right panel).
A General Pattern of Polymorphism Formation
and Collapse
We find that establishment of this polymorphism follows
a characteristic sequence of phases.
Convergence. During phase 1 (Fig. 1, left panel; generations 0 to 10,000), the evolving population simply converges
to the branching point through the gradual adjustment of
phenotypic effects, without any significant between-locus or
within-locus variation being built up. Phase 1 sets the stage
for the establishment of the later polymorphism—by bringing
about a regime of frequency-dependent disruptive selection—
without yet itself contributing to that process.
Symmetric divergence. In phase 2, which commences
right after branching (Fig. 1, middle panel; generations
10,000 to 30,000), the phenotypic differentiation between
alleles grows gradually, due to mutations and allelic substitutions. Closer inspection reveals that all loci become polymorphic during this phase. In particular, we observe two
equally frequent, distinct classes of alleles with equal but
opposite phenotypic effects at each locus. Moreover, the differences between the phenotypic effects of these classes of
alleles are roughly equal for all loci. Consequently, the phenotype segregates as if it were determined by L additive,
EVOLUTIONARY BRANCHING OF MULTILOCUS TRAITS
2229
FIG. 1. Evolution in individual-based model. The three panels of this figure show the distribution of phenotypes during the rapid
convergence to the evolutionary branching point (left), the subsequent phase of diversification at several loci (middle), and the final
phase of evolution at a single locus (right). Small insets A–D show the frequency distribution (frequency is on the vertical axes) of
phenotypes (on the horizontal axes) at four moments during the simulation (indicated by dashed lines). We also ran simulations with
more than three loci: inset E shows an example for L ⫽ 100 loci (time, on the horizontal axis, extends to 500,000 generations; phenotype,
on the vertical axis, ranges from ⫺5.0 to 5.0). Grayscales in the main figure and inset E indicate the frequency of phenotypes. At any
moment in time, the most common phenotype is shown in black, while less common phenotypes are shown in lighter shades of gray.
Note the different scales of the time axis in the three panels. Parameters as listed in the text.
diallelic, diploid loci, with identical pairs of alleles at all loci.
Such genetic systems can give rise to 2L ⫹ 1 different phenotypes, exactly the number of phenotypic classes we observe
in our simulations (Fig. 1, inset A; L ⫽ 3 loci implies seven
such classes).
Between-locus symmetry breaking. In phase 3 (Fig. 1,
middle panel; generations 30,000 to 125,000), phenotypic
variation continues to increase until the distribution of realized phenotypes approximately covers the range from ⫺␮
to ␮. However, the symmetry between loci is broken during
this phase. At some loci, the alleles continue to diversify,
whereas at other loci the differentiation between alleles decreases or alleles are lost altogether (Fig. 1, inset B). Eventually, only one polymorphic locus remains. This effect was
observed to occur independently of the number of loci encoding the ecological strategy (see Fig. 1, inset E, for a simulation with L ⫽ 100) and independently of the parameters
␮ and ␴, as long as ␮ ⬎ ␴. At the remaining polymorphic
locus, two classes of alleles give rise to three distinct classes
of phenotypes (two homozygotes and a heterozygote; Fig. 1,
inset D).
Within-locus symmetry breaking. During phase 4 (Fig. 1,
generations 125,000⫹), phenotypic effects and frequencies
at the last polymorphic locus become asymmetric. This process has previously been studied by Kisdi and Geritz (1999),
who showed that within-locus asymmetries evolve under a
wide range of parameters. During phase 2, and essentially
also during phase 3, the distinct classes of alleles at each
particular locus have equal frequencies and opposite but equal
effects on the phenotype, such that heterozygotes have phe-
notypic effects close to zero. During phase 4, this symmetry
is lost, such that the heterozygote matches one of the two
locally optimal phenotypes, with the other locally optimal
phenotypes being matched by one of the homozygotes (z 艐
␮ and z 艐 ⫺␮; Fig. 1, right panel). The remaining homozygote expresses a poorly adapted phenotype (z 艐 ⫺3␮; Fig.
1, right panel). This makes it evident that the alleles carried
by the latter homozygote (let us refer to its genotype as aa)
have a larger phenotypic effect (xa 艐 ⫺3␮/2) than the alleles
carried by the former homozygote (genotype AA; xA 艐
⫺␮/2). The frequency of the allele a approaches ¼. This is
because, at the time of mating, half of the population consists
of individuals from the first habitat, where only AA individuals stand a fair chance of survival, while the other half
consists of individuals from the second habitat, where only
Aa individuals survive. In Figure 1, the asymmetry in the
phenotype distribution primarily grows during phase 4 (inset
D), but is already initiated to some slight extent during phase
3 (inset C). Beyond these final adjustments, the population’s
phenotypic and allelic composition remains stable.
Replicate simulations for the same set of parameters (at
least 20 replicates per parameter condition) show no variation
on the four-phase pattern described above. Also quantitatively, Figure 1 gives a representative impression of the timing of the different phases. We do observe some variation
between replicates in the length of the phase of symmetric
divergence, though. In about 10% of the replicates, betweenlocus symmetry breaking occurs at a premature stage, such
that polymorphism may already be lost at one locus (or, oc-
2230
G. S.
VAN
DOORN AND U. DIECKMANN
casionally, at two loci) before phenotypic diversification has
come to an end.
The parameters used for the simulations presented in Figure 1 lead to mutational heritabilities, h2m ⫽ 2Lm␴2m , of 6 ⫻
10⫺7 (main figure) and 2 ⫻ 10⫺5 (inset E), and genomic
mutation rates, U ⫽ 2Lm, of 6 ⫻ 10⫺4 (main figure) and 2
⫻ 10⫺2 (inset E). These values span the lower half of the
estimated range of naturally realized values (e.g., Rifkin et
al. 2005). For low values of h2m and U it takes considerable
time until polygenic variation is lost. The whole evolutionary
process occurs more rapidly when the mutational heritability
is enlarged through an increase of the mutation rate, m or of
the variance of mutational effects, ␴2m . However, more pronounced mutations make it more difficult to single out for
further investigation the selection-driven component of evolutionary change; for that reason, we have typically assumed
small values for m and ␴2m (but see Fig. 7).
As we will demonstrate below, the four-phase pattern illustrated in Figure 1 is robustly observed in several variations
of our basic model. While phases 1 and 4 already occur in
single-locus models (Kisdi and Geritz 1999), in this paper
we focus on the new patterns resulting from the symmetry
breaking between loci during phase 3, and thus on processes
that are unique to multilocus models.
DETERMINISTIC MODEL
Derivation of Deterministic Dynamics
We further investigate the observed loss of polymorphism
at all but one locus by analyzing a deterministic approximation of our model. For this purpose we derived deterministic equations for the expected rate of evolutionary change
in allelic effects, assuming that mutations are rare and their
incremental effects are small. Directional evolution then proceeds by steps involving allelic mutation, invasion, and fixation (Metz et al. 1992, 1996; Dieckmann and Law 1996;
Weissing 1996; Hofbauer and Sigmund 1998; Geritz et al.
2002). The outcome of a single step in this process, that is,
whether or not a new mutant allele will be able to invade
and substitute an existing resident allele, is determined by
the invasion fitness of the mutant allele, that is, by the rate
at which the frequency of the mutant allele increases when
it is still rare (Metz et al. 1992, 1996). In a multilocus context,
this quantity will depend on a combination of fitness effects
of the mutant allele in the different genetic backgrounds created by other polymorphic loci (see appendix). Mutant alleles
with positive invasion fitness have a chance to invade the
resident population, and once they have overcome the threat
of accidental extinction by demographic stochasticity (Dieckmann and Law, 1996; Metz et al. 1996), they will go to
fixation (except under certain special and well-understood
circumstances; Geritz et al. 2002). It can be shown that series
of such substitution events result in gradual evolutionary
change at a rate and in a direction that is related to the gradient
of invasion fitness (Dieckmann and Law 1996).
We followed standard procedures for the derivation of invasion fitness and for the subsequent derivation of dynamical
equations for the evolutionary rate of changes in allelic effects (Dieckmann and Law 1996; Kisdi and Geritz 1999;
details are provided in the appendix).
FIG. 2. Evolution in deterministic model. The deterministic approximation of our model tracks the phenotypic differentiation of
alleles at two polymorphic loci. With two alleles at each locus (A
and a at the first, B and b at the second locus), at most nine different
classes of genotypes (indicated by the labels AABB, . . . , aabb) are
present within the population at any moment in time. Individuals
within the same class of genotypes have identical phenotypes. The
phenotypes associated with each class of genotypes and their frequencies change over time, due to evolutionary change in the phenotypic effects of alleles. The time scale of this process may vary
with parameters such as the mutation rate, the mutational variance,
and the population size (see appendix). Parameters as in Figure 1.
Illustration of Deterministic Dynamics
Numerical results for the deterministic model are shown
in Figure 2. The simulation starts with a population located
at the evolutionary branching point, just after a dimorphism
has arisen at two loci. There are two alleles at the first locus,
which we will refer to as A and a, and two alleles at the
second locus, henceforth referred to as B and b (this does not
imply that the alleles A and B are dominant; as before, alleles
act additively on the phenotype).
Until about 1.0 ⫻ 105 generations, the phenotypic effects
of the alleles at both loci diversify rapidly and symmetrically
(corresponding to phase 2 as described above), giving rise
to five phenotypic classes. The difference between the phenotypic effects of alleles B and b then diminishes gradually,
until the allele B is suddenly lost at about 3.5 ⫻ 105 generations (phase 3), so that only three phenotypic classes remain (which of the two loci loses its dimorphism depends
on arbitrarily small initial asymmetries between them). The
difference between the phenotypic effects of alleles A and a
continues to grow throughout phase 3. Finally (phase 4), the
alleles at this locus evolve in such a way that one homozygote
(AA) and the heterozygote (Aa) match the optimal phenotypes, whereas the remaining homozygote (aa) expresses a
suboptimal phenotype. The frequency of the allele a then
EVOLUTIONARY BRANCHING OF MULTILOCUS TRAITS
2231
declines to approximately 0.25. This is again explained by
the fact that 50% of the mating population, that is, the part
contributed by one habitat, will consist mainly of AA individuals, whereas the remaining 50% from the other habitat
will mainly consist of Aa individuals. The alternative outcome also is possible, with the matches provided by aa and
Aa instead (which of these two outcomes is realized depends
on arbitrarily small initial asymmetries between the allelic
effects at the remaining dimorphic locus).
Asymmetries in the initial conditions determine on what
time scale symmetry breaking within and between loci will
occur (asymmetries develop faster when the initial asymmetries are larger). Taking into account the expected value
of initial asymmetries between alleles in the individual-based
simulations, we find good quantitative agreement between
both implementations of our model. Therefore, we use the
deterministic model for further investigation.
Comprehensive Analysis of Deterministic Dynamics
A comprehensive picture of the evolutionary dynamics of
our model can be obtained by focusing on two-locus diallelic
genetics (such as illustrated in Fig. 2) to study the underlying
dynamics in allele space. Let us therefore denote the phenotypic effects of alleles A, a, B, and b as xA, xa, xB, and xb,
respectively. Without loss of generality, we may define
x A ⫽ x̄ ⫹ ␦1 ⫹ ⌬,
(3a)
x a ⫽ x̄ ⫺ ␦1 ⫹ ⌬,
(3b)
x B ⫽ x̄ ⫹ ␦2 ⫺ ⌬,
x b ⫽ x̄ ⫺ ␦2 ⫺ ⌬,
and
(3c)
(3d)
such that x̄ represents the average phenotypic effect of all
four alleles, and ␦1 and ␦2 measure the phenotypic differentiation between alleles at the first and second locus, respectively. The quantities x̄ ⫹ ⌬ and x̄ ⫺ ⌬ then represent
the average phenotypic effects of the alleles at the first and
second locus, respectively. The variable x̄ is indicative of the
asymmetry between alleles at polymorphic loci, whereas the
difference ␦1 ⫺ ␦2 relates to the asymmetry between loci.
Because alleles interact additively within and between loci,
the coefficient ⌬ has no effect at the phenotypic level, and
hence is not subject to selection. This allows us to represent
allele space in three dimensions.
Figure 3 illustrates the different equilibria we find in allele
space. Starting from a population that is monomorphic at
both loci (␦1 ⫽ ␦2 ⫽ 0), evolution first converges to the
evolutionary branching point (indicated as BP in Fig. 3). Any
slight degree of dimorphism developing right at the branching
point (or, alternatively, having been present initially), takes
the population away from this point, toward an equilibrium
at which a symmetric allelic dimorphism is established at
both loci (equilibrium S2). This equilibrium is not stable,
however. Further evolution proceeds toward an equilibrium
at which only one locus supports a symmetric allelic dimorphism (equilibrium S1). Unless the difference between the
patch optima is very large relative to the selection parameter
␴ (for the exact conditions, see Kisdi and Geritz 1999), this
equilibrium is also not stable, such that the final phase of
FIG. 3. Evolution in allele space. Simulations of the deterministic
approximation of our model, started from various initial conditions,
are represented as trajectories in allele space (black lines with arrows). The location of equilibria is indicated by gray circles. The
location of equilibria and their stability properties were calculated
numerically. The thick gray trajectory highlights how evolution
proceeds toward the equilibrium A1 via the equilibria BP, S2, and
S1 (for details see the main text). Notice that in this depiction
trajectories may intersect with other trajectories, because there exist
multiple population genetic equilibria for some combinations of
alleles. Parameters as in Figure 1.
evolution involves the transition to an asymmetric allelic
dimorphism at a single locus (equilibrium A1).
The sequential approach of an initial condition IC toward
the equilibria BP, S2, S1, and A1 in Figure 3 can be recognized in the four different phases of the individual-based
dynamics shown in Figure 1: IC → BP (phase 1), BP → S2
(phase 2), S2 → S1 (phase 3), S1 → A1 (phase 4). The four
different phases are the more pronounced the closer trajectories stay to the itinerary IC → BP → S2 → S1 → A1 (see
Fig. 3). Technically speaking, equilibria like BP, S1, and S2
are called saddle points. Such points are notorious for slowing
down dynamics when being approached closely.
There are several reasons why such approaches dominate
the dynamics of our system. First, due to combinatorial reasons, it is unlikely that only a single locus is polymorphic
shortly after branching. As long as mutations have small
phenotypic effects, one expects the polymorphism to grow
initially at the same rate at every locus. To see why this is
so, suppose that the initial phase of phenotypic diversification
requires n mutations. It is much more likely that these mutations are more or less uniformly distributed over loci than
that all n mutations occurred at the same locus. As long as
n is large relative to the number of loci on which the ecological trait is based, it is therefore probable that the initial
asymmetry between loci is small. This confines trajectories
ejected from the branching point to the plane ␦1 ⫽ ␦2 (Fig.
3). Because mainly combinatorial effects determine the expected direction of evolution from the branching point, de-
2232
G. S.
VAN
DOORN AND U. DIECKMANN
tails of the mutation process can have some impact on the
phase of symmetric divergence. Divergence will typically not
occur at the same rate at every locus, if, for example, the
mutation rate or the mutation step size is higher for some
loci than for others. Similarly, convergence toward the plane
␦1 ⫽ ␦2 may either be supported or hindered by nonlinearities
in the genotype-phenotype mapping.
Second, selection initially tends to decrease the average
phenotypic effect of alleles, 円x̄円, thus selecting for symmetric
(i.e., equal but opposite) phenotypic effects. This effect is a
remnant of the regime of directional selection that drove the
monomorphic population toward the evolutionary branching
point: around ␦1 ⫽ ␦2 ⫽ 0, selection points toward x ⫽ 0
(Fig. 3). In conjunction with the first effect, this means that
trajectories are ejected from the branching point in the direction ␦1 ⫽ ␦2, x̄ ⫽ 0, that is, right toward the equilibrium S2.
Finally, the closer trajectories pass by S2, the closer they
will pass by S1. Because this is a derived effect, the transition
from phase 3 to phase 4 will usually be less sharp than that
from phase 2 to phase 3 (see Fig. 1).
Populations are thus expected to spend considerable time
in the vicinity of the unstable equilibria S2 and S1. This
prediction is corroborated by the individual-based simulation
shown in Figure 1.
ROBUSTNESS
WITH
RESPECT
TO
GENETIC ASSUMPTIONS
So far, we have investigated evolution under frequencydependent disruptive selection in an idealized genetic system,
characterized by free recombination and additive interactions
within and between loci. In addition, we have assumed that
individual mutations have small phenotypic effects. These
simplifying genetic assumptions are habitually made in phenotypic models of evolution, where the details of the underlying genetics are considered to be of secondary importance
(see also Weissing 1996), either because the character under
study is likely to be encoded by many loci or because its
genetic basis is unknown. To overcome these limitations,
below we investigate the robustness of our results with respect to variations of our genetic assumptions.
Genetic Linkage
First, we consider the effects of genetic linkage between
loci. Figure 4 shows numerical results for our deterministic
model with tight linkage between two diallelic loci (the recombination fraction is set to r ⫽ 0.05).
The two loci initially behave as a single locus with four
alleles (given by the haplotypes AB, Ab, aB, and ab). Based
on the results presented in the preceding sections, we expect
that two haplotypes disappear and that the phenotypic effects
of the remaining two haplotypes evolve such that one homozygote and the heterozygote express the two locally optimal phenotypes. This is indeed the case. In the first phase
of the simulation shown in Figure 4 (until about 0.5 ⫻ 106
generations), we observe the emergence of a polymorphism
of five phenotypic classes, but the frequency of two of the
haplotypes (AB and ab) is much higher than that of the other
two haplotypes (Ab and aB). This can be inferred from the
fact that the frequency of the genotypes AAbb and aaBB is
much lower than that of the genotype AaBb. After this initial
FIG. 4. Evolution with tight linkage between loci. Two tightly
linked loci (r ⫽ 0.05), each with two alleles, behave much like a
single locus with four alleles (i.e., combinations of alleles). Consequently, differentiation between loci occurs more slowly than
differentiation between combinations of alleles. Parameters as in
Figure 1, except r ⫽ 0.05.
phase, the phenotypic effect of haplotype ab becomes strongly negative, allowing the homozygote AABB and the heterozygote AaBb to express the two locally optimal phenotypes.
Due to the tight linkage, asymmetries between the loci evolve
more slowly than asymmetries between haplotypes. Eventually however, the polymorphism at one of the loci is lost.
In Figure 4, the allele b disappears shortly after 1.5 ⫻ 106
generations.
These results suggest that linkage between loci does affect
the relative rates at which asymmetries within and between
loci develop, but does not change the partitioning of the
evolutionary dynamics into distinguishable phases, the loss
of polymorphism at all but one locus, and the final pattern
of the evolutionary outcome.
Nonadditive interactions
We also consider the effects of nonadditive interactions
between alleles and between loci. We could relax our assumption of additive genetics by simply imposing fixed, nonadditive interactions (e.g., antagonistic or synergistic interactions). We consider this option less than ideal, because it
would still constrain the evolutionary process. Instead, we
allow for evolutionary change in dominance-recessivity relations and in the relative impacts of the different loci on the
phenotype (we will refer to these relative impacts as the
weights of individual loci). In this extended approach, the
extent to which alleles and loci contribute to the phenotype
is flexible and can be shaped by evolution.
Following the modeling framework introduced by Van
2233
EVOLUTIONARY BRANCHING OF MULTILOCUS TRAITS
TABLE 1. Dependence of the phenotype on allelic effects, allelic parameters (dominance), and modifier loci for the model variant with
nonadditive interactions.
Locus 2
Locus 1
Allele 1
Phenotypic effect
Allelic parameter
Weight of alleles
Alleles at modifier locus
Weight of loci
Phenotype
x11
u11
U11 ⫽ u11/(u11 ⫹ u12)
w11
W1 ⫽ w11 ⫹ w12/(w11
Allele 2
x12
u12
U12 ⫽ u12/(u11 ⫹ u12)
w12
⫹ w12 ⫹ w21 ⫹ w22)
z ⫽ W1 (U11x11 ⫹ U12x12)
Dooren (1999), we implemented this flexibility by assuming
that an individual’s phenotype is determined by the phenotypic effects of the alleles it carries (more precisely, the gene
products of the alleles) and by so-called allelic parameters,
which determine the extent to which the alleles are expressed,
much like regulatory elements in the promotor region of a
gene. In addition, we consider modifier loci (e.g., loci coding
for transcription factors) that affect the level of expression
of all alleles at a given locus. Dominance interactions between alleles derive from the allelic parameters, whereas the
weights of individual loci derive from the expression patterns
at the modifier loci. An allele’s contribution to the phenotype
now depends on its weight relative to the weight of the other
Allele 1
x21
u21
U21 ⫽ u21/(u21 ⫹ u22)
w21
W2 ⫽ w21 ⫹ w22/(w11
⫹ W2 (U21x21 ⫹ U22x22)
Allele 2
x22
u22
U22 ⫽ u22/(u21 ⫹ u22)
w22
⫹ w12 ⫹ w21 ⫹ w22)
allele on the same locus and on the weight of the locus relative
to the weights of the other loci. This is illustrated in Table
1 for a specific example with two loci. Our approach can
easily be extended to allow also for complex epistatic interactions between loci; for the sake of conciseness, we refrain
from illustrating this here. We allowed both the phenotypic
effects of alleles and the allelic parameters to evolve through
mutations with small incremental effects. In addition, we
allowed the weights of loci to evolve through mutations
(again with small incremental effects) of the alleles at modifier loci (one modifier locus for each ecological trait locus).
We assumed free recombination between all loci.
Figure 5 shows numerical results for the extended indi-
FIG. 5. Evolution with variable weights for alleles and loci. Three panels show the distribution of phenotypes in an individual-based
simulation as in Figure 1. The insets A–C, however, do not show frequency distributions, but the average relative weights of loci and
alleles (i.e., the extent to which an allele at a specific locus contributes to the phenotype), at three moments during the simulation
(indicated by dashed lines). The height of the bars represents the weight of a locus (in this simulation we kept track of four loci). For
polymorphic loci, bars consist of a black and white part, indicating the weights of the different alleles that occur at this locus. Gray bars
are used for monomorphic loci. Parameters as in Figure 1, except L ⫽ 4.
2234
G. S.
VAN
DOORN AND U. DIECKMANN
FIG. 6. Evolution with large mutational steps. Even when the whole evolutionary process is reduced to only a small number of allele
substitution events, its predicted phasing is still recognizable. Parameters as in Figure 1, except ␴2m ⫽ 0.1 and m ⫽ 10⫺6.
vidual-based model. We again observe rapid convergence to
the evolutionary branching point, followed by a phase of
phenotypic diversification. Initially, three of four loci become
polymorphic, but eventually only one polymorphic locus remains. Insets A–C in Figure 5 show the relative weights (on
the vertical axis) of the four different loci (on the horizontal
axis) at three moments during the simulation. Gray bars are
used for monomorphic loci; black and white bars are used
for polymorphic loci. The subdivision in a white and black
part represents the relative weights of the two different alleles
that occur on a polymorphic locus. During the initial phase
of differentiation (corresponding to phase 2 as described
above), the alleles at all polymorphic loci diversify symmetrically (Fig. 5, inset A, at 3.0 ⫻ 104 generations). As long
as selection favors further diversification, there is directional
selection to increase the weight of polymorphic loci, which
contribute to population-level phenotypic diversity, relative
to the weight of the one monomorphic locus (locus 2), which
does not. At this time, selection on the allelic parameters
(i.e., on dominance) is still virtually absent. Later, however,
the asymmetries between loci grow (corresponding to phase
3 as described above), until only one polymorphic locus remains (locus 1). During this phase, selection on the relative
weights of polymorphic loci is disruptive and acts alongside
selection on allelic effects (which is stabilizing for some loci
but diversifying for other loci), such that the locus with the
largest differentiation between alleles eventually contributes
to the phenotype with the largest relative weight (Fig. 5, inset
B, at 8.0 ⫻ 104 generations). All along, the interaction between alleles at a single locus has remained additive, that is,
the alleles at polymorphic loci have equal relative weights.
However, selection for dominance-recessivity interactions
between alleles arises as soon as asymmetries evolve between
alleles at the remaining polymorphic locus (corresponding to
phase 4 as described above). The relative weight of one of
the alleles increases, such that, eventually, the phenotype of
the (otherwise) maladapted homozygote coincides with the
locally optimal phenotype matched by the heterozygote (Fig.
5, inset C, at 7.0 ⫻ 105 generations).
These results show that the evolution of nonadditive interactions between alleles and the evolution of locus weights
are expected to act alongside the evolution of allelic effects,
representing alternative pathways along which the symmetry
between and within loci can be broken. The relative contributions of the evolution of allelic effects (the evolution of
the gene products) versus the evolution of the weights of
alleles and loci (the evolution of gene regulation) will depend
on factors like the relative mutation rates of the phenotypic
effects, the allelic parameters, and the modifier alleles. All
key predictions of our preceding analysis are corroborated
even in this extended model. In particular, the characteristic
phasing of dynamics from the initial diversification to the
final outcome is robustly recovered.
Large Mutations
As a third check on the robustness of our results, we explore the effects of large mutational step sizes. Figure 6 shows
numerical results for our original individual-based model,
with all parameters except the mutational step size ␴m and
the mutation rate m chosen exactly as in Figure 1. In Figure
6, the variance of mutational effects ␴2m was set 100 times
larger than in Figure 1, and the mutation rate was set 100
times smaller, such that the expected rate of directional evolution, which scales with m␴2m (Dieckmann and Law 1996;
see also the appendix), was identical for both simulations.
These results show that with large mutational steps the
whole evolutionary process—of convergence to the branching point, loss of polymorphism on all-but-one locus, and
asymmetric differentiation of alleles at the remaining polymorphic locus—is reduced to only a small number of allele
substitution events (which can be recognized individually as
discontinuities in Fig. 6). Consequently, the stochasticity of
the mutation process is much more pronounced, and the var-
EVOLUTIONARY BRANCHING OF MULTILOCUS TRAITS
iation between replicate simulations is larger. Yet, the average behavior of replicates does not deviate qualitatively
from the predictions of our deterministic model. Also the
diminished phasing of the evolutionary process is just as
predicted, because smaller mutational steps make it easier
for the genetic system to track the saddle connections that
lie at the heart of the process. Figure 6 shows that we can
still recognize the different phases discussed before, even
when mutation effects are not small. It is clear that the weaker
selection and the larger the mutational step size, the more
strongly the stochasticity of the mutation process will blur
the selection-driven, deterministic component of evolutionary change.
DISCUSSION
Our results show that frequency-dependent disruptive selection is less powerful in maintaining polygenic variation
than one would naively expect. Frequency-dependent disruptive selection does not lead to the establishment of genetic
polymorphism at a large number of loci. Instead, genetic
variation is concentrated at a single locus with large phenotypic effect. We observed this outcome in individual-based
simulations and in an analytical model, under a range of
genetic assumptions, which gives confidence in the robustness of the results. The identified pattern of polymorphism
formation and collapse is likely to be widely applicable.
The dynamics observed in our model suggest a conceptual
link between the different effects of frequency-dependent disruptive selection observed in quantitative genetics and adaptive dynamics models. In the initial phase of diversification,
all loci are polymorphic, and the phenotypic differentiation
of alleles at each locus is small. Hence, a large number of
loci contribute to the genetic variation, and each locus has a
small effect on the phenotype. Not surprisingly, the dynamics
shortly after evolutionary branching therefore much resembles the maintenance of variation as observed in quantitative
genetics models, where disruptive selection leads to the gradual broadening of a continuous phenotype distribution. Eventually, however, genetic variation becomes concentrated at a
single locus, which contributes increasingly strongly to phenotypic variation. In this situation quantitative genetics methods become inaccurate: we observe the emergence of discrete
clusters of phenotypes that create a situation better analyzed
by adaptive dynamics methods or by classical population
genetics.
The phenomena of polymorphism formation and collapse
observed in our model are a straightforward consequence of
the fact that frequency-dependent selection generates a dynamic selection regime. It is a defining feature of frequency
dependence that the intensity and direction of selection
changes as evolution proceeds, a consequence of the feedback
between a population and its environment. In the context of
our model, the population first experiences directional selection toward the evolutionary branching point, then disruptive
selection at the branching point (leading to diversification at
all loci), and subsequently again a type of stabilizing selection (favoring two discrete phenotypes at the patch optima).
Selection turns from disruptive to stabilizing as soon as
the phenotypic variation in the population has become large
2235
enough for the optimal phenotypes in the two patches to occur
at appreciable frequencies. At that point, there is no further
selection for diversification. Yet, intermediate phenotypes remain at a selective disadvantage. It is an unavoidable consequence of sexual reproduction—at least as long as individuals mate at random—that such intermediate phenotypes
are generated, but, for combinatorial reasons, their frequency
is lowest when all genetic variation is concentrated at a single
locus. This explains why all loci, except one, eventually become monomorphic. Subsequent evolution, involving symmetry breaking between alleles at the remaining polymorphic
locus, increases population mean fitness by further reducing
the frequency of maladapted individuals.
Although here we have analyzed only Levene’s soft-selection model, we expect that our conclusions apply to a
broad class of systems subject to frequency-dependent disruptive selection. Adaptive dynamics theory has revealed the
generic shape of fitness landscapes around evolutionary
branching points (e.g., Geritz et al. 1997), and adaptive dynamics models have shown that such branching points can
be created by a plethora of different ecological mechanisms,
including all three fundamental types of ecological interaction (e.g., Doebeli and Dieckmann 2000). In particular, we
expect to observe similar evolutionary phenomena in all cases
where the coexistence of an arbitrarily large number of replicators is precluded by a competitive exclusion principle
(Gyllenberg and Meszéna 2005). Such systems must, at some
level of diversity, exhibit a transition from disruptive to stabilizing selection favoring the evolution of a discrete, limited
set of phenotypes. In our Levene-type model, the number of
coexisting replicators is bounded by the number of different
habitats, and this sets an upper limit on the number of loci
expected to remain polymorphic in long-term evolution. In
a two-niche environment, at most one locus remains polymorphic. In environments with more than two habitats, polymorphism might be maintained at more than one locus (or
more that two alleles might segregate at a single locus), but
the number of polymorphic loci is always smaller than the
number of habitat types (data available from the authors).
In a somewhat different context, Spichtig and Kawecki
(2004), who recently also analyzed a multilocus version of
Levene’s model, came to a conclusion similar to ours. While
their analysis addressed the dynamics and the equilibrium
frequencies of a fixed set of alleles, other aspects of the two
models are similar, allowing for a detailed comparison of
results. Spichtig and Kawecki (2004) argued that the capacity
of soft selection to maintain polygenic variation is smaller
than one would expect based on single-locus models. Their
conclusion, however, applies to parameter regimes for which
evolutionary branching does not occur because the fitness of
intermediate phenotypes is high. Under these conditions,
polygenic variation is not maintained, due to the fact that the
average phenotype of a polygenic character can be accurately
matched with the optimal phenotype without requiring a polymorphism of alleles at individual loci (i.e., with all loci being
homozygous, and, hence, with the population being monomorphic). This conclusion does not apply to a single locus,
where the realization of an intermediary phenotype typically
requires a heterozygous genotype (and, hence, a polymorphic
population). Unlike for single-locus characters, the variation
2236
G. S.
VAN
DOORN AND U. DIECKMANN
of polygenic characters can therefore be low, irrespective of
the mean phenotype.
In contrast, our conclusion applies to the maintenance of
polygenic variation after evolutionary branching; that is, it
concerns a complementary parameter regime. In this case,
the explanation for the loss of polygenic variation is different
and stems from the fact that a single-locus polymorphism
allows for a maximal level of phenotypic variation: given a
certain degree of overall differentiation between alleles, the
phenotypic variance in the population is highest when the
polymorphism is concentrated at a single locus. Under conditions that allow for evolutionary branching, a polymorphism of differentiated alleles at a single locus is therefore
the most favorable configuration that can be attained within
the limits set by sexual reproduction. It allows for the lowest
possible frequency of the intermediate phenotypes that are
at selective disadvantage in the parameter regime considered
here.
Obviously, a single-locus polymorphism will only be favored over polygenic variation when the phenotypic effects
of individual alleles are considerable, such that a polymorphism at a single locus can give rise to substantial phenotypic
variation. In our model, the phenotypic effects of individual
alleles can become arbitrarily large, as a cumulative result
of many mutations with small phenotypic effects. In models
that do not incorporate mutation, where the set of alleles is
kept fixed and the phenotypic effects of individual alleles are
limited, a polymorphism of alleles at a single locus can only
give rise to a modest level of phenotypic variation. In such
a situation, we would expect variation to be maintained at
multiple loci, because this is the only way to maintain sufficient genetic variation (an expectation confirmed by Bürger
2002a,b; Spichtig and Kawecki 2004).
This highlights another contrast between our analysis and
studies of frequency-dependent disruptive selection that investigate allele-frequency changes and the stability properties
of population genetic equilibria of a predefined set of alleles.
The latter yield conditions for the short-term maintenance of
genetic variation but do not provide insights about long-term
evolution, which occurs through the substitution of the existing alleles by novel, mutant alleles (Eshel 1996). This process is explicitly considered in our model. However, we have
largely neglected potential constraints on the evolution of
allelic effects; in the absence of knowledge that warrants
more specific assumptions, we have merely assumed the mutational step size to be small. The mechanistic details of the
mutation process, the development of the phenotype, and so
on will be important to calibrate the different time scales of
short-term and long-term evolution relative to one another.
Such calibration is necessary to interpret observed patterns
of polymorphism of quantitative traits in empirical systems.
Conversely, a detailed comparison of these patterns to the
evolutionary predictions made by short- and long-term evolution models may give insights in the importance and nature
of evolutionary constraints on polygenic variation.
ACKNOWLEDGMENTS
GSvD thanks F. Weissing for useful suggestions and gratefully acknowledges support from the International Institute
for Applied Systems Analysis (IIASA) and the Center for
Ecological and Evolutionary Studies (CEES) enabling several
visits to IIASA. UD gratefully acknowledges financial support from the Austrian Science Fund; the Austrian Federal
Ministry of Education, Science, and Cultural Affairs; and the
European Research Training Network ModLife (Modern LifeHistory Theory and its Application to the Management of
Natural Resources), funded through the Human Potential Programme of the European Commission. The authors thank T.
F. Hansen, R. Bürger, M. Kopp, and an anonymous reviewer
for their helpful comments on an earlier version of the manuscript.
LITERATURE CITED
Abrams, P. A. 2001. Modeling the adaptive dynamics of traits involved in inter- and intraspecific competition: an assessment of
three methods. Ecol. Lett. 4:166–175.
Bulmer, M. 1980. The mathematical theory of quantitative genetics.
Clarendon Press, Oxford, U.K.
Bürger, R. 2000. The mathematical theory of selection, recombination and Mutation. John Wiley and Sons, Chicester, U.K.
———. 2002a. Additive genetic variation under intraspecific competition and stabilizing selection. J. Theor. Popul. Biol. 61:
197–213.
———. 2002b. On a genetic model of intraspecific competition and
stabilizing selection. Am. Nat. 160:661–682.
Bürger, R., G. P. Wagner, and F. Stettinger. 1989. How much heritable variation can be maintained in finite populations by mutation-selection balance? Evolution 43:1748–1766.
Caswell, H. 1989. Matrix population models. Sinauer Associates,
Sunderland, MA.
Clarke, B. 1972. Density-dependent selection. Am. Nat. 106:1–13.
Cockerham, C. C., P. M. Burrows, S. S. Young, and T. Prout. 1972.
Frequency-dependent selection in randomly mating populations.
Am. Nat. 106:493–515.
Cressman, R. 1992. The stability concept of evolutionary game
theory: a dynamic approach. Springer-Verlag, Berlin.
Dieckmann, U, and R. Law. 1996. The dynamical theory of coevolution: a derivation from stochastic ecological processes. J.
Math. Biol. 34:579–612.
Doebeli, M, and U. Dieckmann. 2000. Evolutionary branching and
sympatric speciation caused by different types of ecological interactions. Am. Nat. 156:S77–S101.
Eshel, I. 1996. On the changing concept of evolutionary population
stability as a reflection of a changing point of view in the quantitative theory of evolution. J. Math. Biol. 34:485–510.
Felsenstein, J. 1976. The theoretical population genetics of variable
selection and migration. Annu. Rev. Genet. 10:253–280.
Fisher, R. A. 1930. The genetical theory of natural selection. Oxford
Univ. Press, Oxford, U.K.
Geritz, S. A. H., J. A. J. Metz, É. Kisdi, and G. Meszéna. 1997.
Dynamics of adaptation and evolutionary branching. Phys. Rev.
Lett. 78:2024–2027.
Geritz, S. A. H., É. Kisdi, G. Meszéna, and J. A. J. Metz. 1998.
Evolutionarily singular strategies and the adaptive growth and
branching of the evolutionary tree. Evol. Ecol. 12:35–57.
Geritz, S. A. H., M. Gyllenberg, F. J. A. Jacobs, and K. Parvinen.
2002. Invasion dynamics and attractor inheritance. J. Math. Biol.
44:548–560.
Gyllenberg, M., and G. Meszéna. 2005. On the impossibility of
coexistence of infinitely many strategies. J. Math. Biol. 50:
133–160.
Hedrick, P. W., M. E. Ginevan, and E. P. Ewing. 1976. Genetic
polymorphism in heterogeneous environments. Annu. Rev. Ecol.
Syst. 7:1–32.
Hofbauer, J., and K. Sigmund. 1998. Evolutionary games and population dynamics. Ch. 9: Adaptive dynamics. Cambridge Univ.
Press, Cambridge, U.K.
Keightley, P. D., and W. G. Hill. 1983. Effects of linkage on re-
2237
EVOLUTIONARY BRANCHING OF MULTILOCUS TRAITS
sponse to directional selection from new mutations. Genet. Res.
42:193–206.
Kisdi, É., and S. A. H. Geritz. 1999. Adaptive dynamics in allele
space: evolution of genetic polymorphism by small mutations
in a heterogeneous environment. Evolution 53:993–1008.
Levene, H. 1953. Genetic equilibrium when more than one ecological niche is available. Am. Nat. 87:331–333.
Lewontin, R. 1958. A general method for investigating the equilibrium of gene frequency in a population. Genetics 43:419–434.
Maynard Smith, J. 1982. Evolution and the theory of games. Cambridge Univ. Press, Cambridge, U.K.
Metz, J. A. J., R. M. Nisbet, and S. A. H. Geritz. 1992. How should
we define ‘‘fitness’’ for general ecological scenarios? Trends
Ecol. Evol. 7:198–202.
Metz, J. A. J., S. A. H. Geritz, G. Meszéna, F. J. A. Jacobs, and J.
S. van Heerwaarden. 1996. Adaptive dynamics, a geometrical
study of the consequences of nearly faithful reproduction. Pp.
183–231 in S. J. van Strien and S. M. Verduyn Lunel, eds.
Stochastic and spatial structures of dynamical systems. Elsevier,
North Holland.
Ravigné, V., I. Olivieri, and U. Dieckmann. 2004. Implications of
habitat choice for protected polymorphisms. Evol. Ecol. Res. 6:
125–145.
Rifkin, S. A., D. Houle, J. Kim, and K. P. White. 2005. A mutation
accumulation assay reveals a broad capacity for rapid evolution
of gene expression. Nature 438:220–223.
Roughgarden, J. 1979. Theory of population genetics and evolutionary ecology: an introduction. MacMillan, New York.
Slatkin, M. 1979. Frequency- and density-dependent selection on
a quantitative character. Genetics 93:755–771.
Spichtig, M., and T. J. Kawecki. 2004. The maintenance (or not)
of polygenic variation by soft selection in heterogeneous environments. Am. Nat. 164:70–84.
Taylor, P. D. 1996. Inclusive fitness arguments in genetic models
of behaviour. J. Math. Biol. 34:654–674.
Van Dooren, T. J. M. 1999. The evolutionary ecology of dominancerecessivity. J. Theor. Biol. 198:519–532.
Weissing, F. J. 1996. Genetic versus phenotypic models of selection: Can genetics be neglected in a long-term perspective? J.
Math. Biol. 34:533–555.
second habitat, and the factor ½ simply reflects the assumption that
half of the individuals in the mating population are recruited from
either habitat.
When the resident population is polymorphic at one or more loci,
the derivation of invasion fitness becomes more complicated, because we then need to keep track of the frequencies of the different
resident genotypes. The mutant allele may then also occur in different genetic backgrounds, consisting of different combinations of
resident alleles. To keep the analysis tractable, we will restrict ourselves here to the simplest interesting case, by considering a resident
population that is polymorphic at two loci (L ⫽ 2). We denote the
alleles at the first locus by A and a, and the alleles at the second
locus by B and b (as mentioned in the main text, this notation does
not imply that the alleles A and B are dominant). The phenotypic
effects of the alleles are denoted by xA, xa, xB, and xb. If xA ⫽ xa
or xB ⫽ xb, the resident population is monomorphic at the corresponding locus. We also consider a mutant allele M, with phenotypic
effect xm, that has arisen through mutation of the allele A at the
first locus (other mutant alleles are dealt with analogously).
We choose to describe the dynamics of the resident and mutant
allele frequencies in terms of the frequencies of the haploid gametes
in which they occur: fg denotes the frequency of the gamete g (g
⫽ AB, Ab, aB, MB, or Mb) in adults at the moment of reproduction,
that is, after viability selection has occurred. We follow the life
cycle of our model to determine its effect on these gamete frequencies.
Random mating. We first compute the frequency Fg⬘ g of the genotype gg⬘ in the offspring before viability selection. Because mating is random, the frequency of offspring carrying the genotype
gg⬘, which arises from the combination of gametes g and g⬘ is given
by the product of the corresponding gamete frequencies in the parents, Fg⬘ g⬘ ⫽ fg fg⬘.
Viability selection. Viability selection changes the genotype frequencies in the offspring, such that the frequency Fgg⬘ of the genotype gg⬘ after viability selection is, similarly to equation (A1),
Corresponding Editor: T. Hansen
where Zgg⬘ denotes the phenotype encoded by the genotype gg⬘ (e.g.,
ZABAb ⫽ 2xA ⫹ xB ⫹ xb and ZabMb ⫽ xa ⫹ xM ⫹ 2xb) and v̄i is the
average viability in habitat i. While the mutant allele is rare, average
viabilities do not depend on the mutant’s genotype frequencies,
APPENDIX: DERIVATION OF DETERMINISTIC APPROXIMATION
Here we derive an analytical deterministic approximation that
captures the dynamics in our individual-based stochastic simulation
model. To enable this complementary treatment, we assume that
mutations occur rarely, such that mutant alleles arise in a resident
population that is close to its population genetic equilibrium.
Consequently, a mutant allele interacts only with the currently
predominant resident alleles, which were successful at ousting previous mutant alleles. We also assume that the population is sufficiently large, such that we may neglect stochasticity in the dynamics
of the frequencies of resident alleles, and that changes in the phenotypic effects of alleles caused by individual mutations are typically small, such that it is meaningful to approximate the long-term
dynamics of phenotypic effects deterministically.
The invasion fitness ␭ specifies the geometric rate of increase of
the abundance of a mutant allele while it is rare (e.g. just after it
has arisen by mutation; Metz et al. 1992, 1996). When a mutant
arises in an otherwise genetically monomorphic resident population,
all resident individuals have the same phenotype ẑ and all individuals that carry a mutant allele have the same phenotype Z. This
greatly simplifies the derivation of invasion fitness in our model
(e.g., Kisdi and Geritz 1999), which under such conditions is given
by
␭⫽
1
[v (z)/v1 (ẑ) ⫹ v2 (z)/v2 (ẑ)].
2 1
(A1)
The first and second term in the square bracket represent, respectively, the relative viabilities of mutant individuals in the first and
1
F gg⬘ ⫽ F ⬘gg⬘ [v1 (z gg⬘ )/v̄1 ⫹ v2 (z gg⬘ )/v̄2 ],
2
v̄ i ⫽
冘
g,g⬘⫽AB,Ab,aB,ab
F ⬘gg⬘ v i (z gg⬘ ).
(A2)
(A3)
Gamete production. After viability selection, the next generation is produced through sexual reproduction. The frequencies of
the different resident gametes are determined straightforwardly
from the resident genotype frequencies after viability selection. For
example,
f AB ⫽ F ABAB ⫹
⫹
1
2
冘
g⫽Ab,aB
(F ABg ⫹ F gAB )
1
1
(1 ⫺ r)(F ABab ⫹ F abAB ) ⫹ r(F AbaB ⫹ F aBAb ),
2
2
(A4)
where r is the coefficient of recombination between the two loci.
The mutant’s genotype frequencies do not appear in equation (A4),
since the frequency of the mutant allele is initially negligible.
Equations (A2) to (A4) define a recurrence relation for the resident gamete frequencies. This recurrence relation can be iterated
until these frequencies converge to a stable equilibrium (reflecting
our assumption that resident populations attain their population genetic equilibrium by the time a mutant arises).
For the mutant gamete frequencies we obtain, analogously to
equation (A4),
2238
G. S.
f MB ⫽
1
2
⫹
冘
g⫽AB,aB
f Mb ⫽
1
2
冘
冘
冘
polymorphic resident background is now given by the geometric
rate of increase of the mutant allele frequency, which equals the
dominant eigenvalue of the matrix F ⫹ (xM ⫺ xA)W. For small
円xM ⫺ xA円 it can be shown (e.g., Caswell 1989; Taylor 1996) that
(F MBg ⫹ F gMB )
g⫽Ab,ab
␭ ⫽ 1 ⫹ (x M ⫺ x A )
(F Mbg ⫹ F gMb ),
g⫽AB,aB
g⫽Ab,ab
DOORN AND U. DIECKMANN
(F MBg ⫹ F gMB )
1
(1 ⫺ r)
2
1
⫹ r
2
VAN
1
(1 ⫺ r)
2
⫹
1
r
2
冘
冘
g⫽AB,aB
g⫽Ab,ab
v ⫽ (F ABAb ⫹ F ABab , F ABAb ⫹ F AbaB )
u⫽
(F Mbg ⫹ F gMb )
(A5)
Here we again use the fact that the mutant allele is rare initially,
which allows us to neglect the frequency of individuals that are
homozygous for the mutant allele.
For mutant alleles M that differ only slightly from the resident
allele A, 円xM ⫺ xA円 is small, and we may use first-order Taylor
expansions to approximate the viabilities of phenotypes affected by
the mutant allele. For example,
[
]
艐 [1 ⫺ (x M ⫺ x A )(z ABg ⫺ ␮ i )/␴ 2 ]v i (z ABg ).
(A6)
Using these approximations and equation (A2), we rewrite the mutant genotype frequencies. For the mutant genotype frequencies
FMBg, for example, this yields
F MBg ⫽
f MB
F
⫹ f MB (x M ⫺ x A )W AB
g ,
f AB ABg
(A7)
where
1
W gg⬘ ⫽ ⫺ f g␴ ⫺2 [(z gg⬘ ⫺ ␮1 )v1 (z gg⬘ )/v̄1
2
⫹ (z gg⬘ ⫺ ␮2 )v2 (z gg⬘ )/v̄2 ].
(A8)
We substitute equation (A7) and analogous expressions for the
genotype frequencies FgMB, FMbg, and FgMb into equations (A5), to
obtain, after some rearrangement, the following recurrence relation
for the change of mutant gamete frequencies from one generation
to the next,
冢 f 冣 → [F ⫹ (x
f MB
M
⫺ x A )W]
Mb
冢 f 冣,
f MB
(A9)
Mb
where the matrices F and W are defined as
F⫽
[
]
1 ⫺ rf ⫺AB1 (F ABAb ⫹ F AbaB )
rf ⫺Ab1 (F ABAb ⫹ F AbaB )
rf ⫺AB1 (F ABAb ⫹ F ABab )
1 ⫺ rf ⫺Ab1 (F ABAb ⫹ F ABab )
⎡
⎢
⎢
冘
g⫽AB,aB
⫹ (1 ⫺ r)
W⫽⎢
⎣
W AB
g
r
冘
冘
g⫽Ab,ab
g⫽Ab,ab
W AB
g
r
冘
g⫽AB,aB
W AB
g
冘
g⫽Ab,ab
,
(A10)
W Ab
g
⎤
⎥
⎥
⫹ (1 ⫺ r)
冘
g⫽AB,aB
冢 f 冣,
f AB
(A13a)
(A13b)
which are the dominant left and right eigenvectors of the matrix F,
respectively. Under suitable assumptions (Dieckmann and Law
1996; Weissing 1996; Hofbauer and Sigmund 1998), the invasion
fitness can be used to describe the long-term rate and direction of
a series of allelic substitution events. Indeed, using equation (A12)
and following the derivation scheme employed by Dieckmann and
Law (1996), it can be shown that the expected evolutionary rate of
change of the phenotypic effect of the currently resident allele A
at the first locus satisfies
dx A
⫽
dt
冕
2Nm( f AB ⫹ f Ab ) · M(x M 円 x A )
[
⫻ ␣ max 0, (x M ⫺ x A )
]
vWu
· (x M ⫺ x A ) dx M ,
vu
(A14)
where t measures evolutionary time in generations. The first factor
in the integrand above is the rate at which new mutant alleles arise:
the frequency of allele A is given by fAB ⫹ fAb, the total number of
alleles in a diploid population of size N is 2N, and m equals the
mutation rate per generation. The second factor is the probability
density according to which a mutation changes the phenotypic effect
at the first locus from xA to xM. The third factor is the probability
that the mutant allele will successfully invade. This probability is
zero when the mutant allele has a geometric rate of increase below
that of the resident allele and otherwise is proportional to the fitness
advantage s of the mutant allele, as long as s is small. This explains
the function ␣ max(0, s), with ␣ denoting the constant of proportionality, and with s ⫽ (xM ⫺ xA)(vWu)/(vu) following from equation
(A12). For offspring numbers varying according to a Poisson distribution, we obtain ␣ ⫽ 2. If the mutant allele succeeds to invade,
this causes a change of the resident allele: away from the evolutionary branching point (and from population dynamical bifurcation
points), successful invasion of the mutant allele implies that it will
eventually replace the resident allele (Geritz et al. 2002). Successful
invasion thus means that the phenotypic effect of the currently
resident allele will change by an amount xM ⫺ xA, which explains
the integrand’s fourth factor.
Collecting all terms that are independent of xM in front of the
integral, and realizing that the integrand above vanishes along half
its range because only mutant alleles with either xM ⬎ xA or xM ⬍
xA can successfully invade, we can rewrite equation (A14) as
dx A
vWu
⫽ 2Nm( f AB ⫹ f Ab )␣
dt
vu
⫻
⎥.
W Ab
g
and
Ab
(F MBg ⫹ F gMB ).
1
v i (z MBg ) ⫽ exp ⫺ (z MBg ⫺ ␮ i ) 2 /␴ 2
2
(A12)
The term vWu/(vu), which represents the fitness gradient, varies
with the vectors
(F Mbg ⫹ F gMb )
⫹
vWu
.
vu
1
2
冕
(x M ⫺ x A ) 2 M(x M 円 x A ) dx M .
(A15)
Denoting the variance of mutational effects by ␴2m, we therefore
finally obtain
W Ab
g
dx A
vWu
⫽ Nm␣␴ 2m ( f AB ⫹ f Ab )
.
dt
vu
⎦
(A11)
The invasion fitness ␭ of the mutant allele in the considered
(A16)
Equations for the rate of change in the phenotypic effects of the
alleles a, B, and b, are derived analogously.